Design systems in the age of generated interfaces

For a decade the case for a design system was speed: reuse the button, ship the screen faster. A model can now produce a plausible screen in seconds. Speed is no longer the scarce thing. Consistency, accessibility and intent are.

That shifts what the system is for. Ours changed in four ways.

Constraints became the product

A generated screen is only as good as the rules it was generated under. We rewrote our tokens and component docs so a model can consume them: explicit spacing scales, named states, and prose that says when a pattern must not be used. The system is now a specification first and a Figma library second.

Fewer components, stricter ones

Every variant is a way for generated output to drift. We cut the component count by a third and locked what remained. A model that can only reach for six layout primitives produces screens that look like ours; one that can reach for sixty produces screens that look like everyone's.

Accessibility moved into the tokens

Contrast pairs, focus styles and minimum target sizes are now properties of the tokens themselves rather than a checklist applied afterwards. Generated interfaces inherit them by construction, which is the only way the check reliably happens.

Review is where the taste lives

Generation is cheap; judgement is not. The designer's time moved from drawing screens to reviewing them against the system and, more importantly, against the problem. That is a better use of the skill, and it is the part no model has taken.